## How the Battle for Control Could Crush AI’s Promise

## Source details

**Canonical URL:** [How the Battle for Control Could Crush AI’s Promise](https://www.imf.org/en/publications/fandd/issues/2025/09/how-the-battle-for-control-could-crush-ais-promise-carl-benedikt-frey)

## Other formats

- [Markdown version](/en/publications/fandd/issues/2025/09/how-the-battle-for-control-could-crush-ais-promise-carl-benedikt-frey/index.md)
- [Structured JSON version](/en/publications/fandd/issues/2025/09/how-the-battle-for-control-could-crush-ais-promise-carl-benedikt-frey/index.json)
- [Bundle manifest](/en/publications/fandd/issues/2025/09/how-the-battle-for-control-could-crush-ais-promise-carl-benedikt-frey/bundle-manifest.json)

## Bibliographic details
- Authors: CARL BENEDIKT FREY
- Published: September 3, 2025

---

### Thesis and central argument
- A shift toward centralization and concentration could snuff out technology’s productive potential.
- True breakthroughs emerge from decentralized exploration and widening the arena of experimentation; centralized scale and concentration risk stagnation.
- Sustaining a policy regime that safeguards competitive arenas, not the fortunes of particular firms, is essential for AI to deliver productivity gains.

### Historical evidence and analogies
- Soviet Union
  - Despite early technological successes (Sputnik, Yuri Gagarin), the USSR collapsed as the computer revolution took off.
  - Zelenograd’s centrally controlled innovation contrasted with Silicon Valley’s decentralized experimentation; institutionally inhospitable environments limited exploration.
  - Friedrich Hayek: central planners lacked essential local knowledge, could manage standardized operations but floundered during technological uncertainty.
- Japan
  - Japanese autoworkers were 17 percent more productive than their US counterparts by 1980.
  - Japan focused on process improvements: Edwin Mansfield found roughly two-thirds of Japanese R&D targeted process improvements.
  - Japanese coordination and keiretsu structures favored incremental refinement over frontier product innovation, contributing to a stall when the center of innovation shifted to software.
- Western Europe
  - Postwar coordinated capitalism supported catch-up growth but became an obstacle when the economy needed to pivot to frontier innovation.
  - France’s indicative planning and Italy’s state-owned enterprises were effective for incremental, predictable progress but ill-suited to rapid, uncertain technological change.
  - Southern European nations experienced prolonged stagnation during the computer revolution, often described as “two lost decades.”
- United States in the computer era
  - US antitrust policy (rooted in the 1890 Sherman Antitrust Act) pried open markets (e.g., unbundling IBM, breaking up AT&T), enabling entrepreneurial dynamism that powered the internet and software-led gains.
  - Regions organized around fierce competition (Silicon Valley) outperformed more hierarchical clusters (Route 128).

### Frontier innovation and AI-specific findings
- Scaling is insufficient for frontier breakthroughs; exploration matters more than perfecting formalized systems.
- Large language models (LLMs)
  - Grew 10,000-fold in scale between 2019 and 2024 yet still scored only about 5 percent on the ARC reasoning benchmark.
  - Leaner approaches such as program search have topped 20 percent on the same benchmark.
  - Newer in-context learning methods are “racing ahead.”
- Embodied knowledge and limitations of centralized models
  - Language models trained on the entire internet still lack sensorimotor experience comparable to any four-year-old.
  - Hans Moravec’s observation: what is effortless for humans (e.g., walking a trail) remains hard for machines, and vice versa.
  - Historical examples illustrate statistical-consensus failure modes: an LLM trained in 1633 would uphold geocentrism; given 19th century literature, it would deny human flight.
  - Demis Hassabis (Google DeepMind) concedes true artificial general intelligence may need “several more innovations.”

### Risks from centralization and concentration today
- China
  - Recentralization: licenses, credit, and contracts favor politically reliable conglomerates; antitrust law is wielded selectively; anti-corruption campaigns make loyalty prerequisite for survival.
  - Provincial experimentation has withered; officials chase crude indicators (e.g., patent counts), flooding registries with low-value filings.
  - Patronage and loyalty are displacing competence, eroding the state’s capacity to nurture frontier-level innovation and pushing growth toward slower, less-innovation-driven paths.
  - Private or foreign-backed firms remain the most dynamic sectors; state-owned enterprises lag.
  - Firms lacking strong political connections (example: DeepSeek) tend to be most innovative but remain vulnerable without robust legal protections.
- United States and incumbent concentration
  - Labor mobility is hampered by a web of noncompete clauses, curbing tacit-knowledge flows and discouraging new firm creation.
  - Incumbent lobbying hard-codes regulatory advantages (patent extensions, sector-specific licensing hurdles) reducing competitive dynamism.
  - Market concentration in AI-related inputs and platforms:
    - Microsoft’s alliance with OpenAI controls about 70 percent of the commercial LLM market.
    - Nvidia provides about 92 percent of the specialized graphics-processing units (GPUs) used to train these models.
  - Alphabet, Amazon, and Meta, alongside Microsoft and Nvidia, have been acquiring stakes in promising AI start-ups, consolidating influence.

### Key statistics and exact figures drawn from the text
- LLM scale increase: 10,000-fold between 2019 and 2024.
- LLM performance: about 5 percent on the ARC reasoning benchmark.
- Program search performance: topped 20 percent (on ARC).
- Human sensorimotor benchmark: any four-year-old (qualitative comparison).
- Historical training examples: 1633; 19th century (qualitative).
- Japanese autoworker productivity advantage: 17 percent by 1980.
- US antitrust origin: 1890 Sherman Antitrust Act (historical reference).
- Market concentration:
  - Microsoft/OpenAI alliance: about 70 percent of the commercial LLM market.
  - Nvidia share of specialized GPUs: about 92 percent.

### Policy implications and recommendations
- Protect and expand competitive arenas rather than protecting incumbent firms:
  - Prevent regulatory capture that hard-codes advantages for incumbents (patent extensions, licensing hurdles).
  - Limit practices that impede labor mobility (e.g., overly broad noncompete clauses) to preserve tacit-knowledge flows essential for startup-driven innovation.
- Lower barriers to entry and widen experimentation:
  - Reduce centralization that channels licenses, credit, and contracts toward politically reliable conglomerates.
  - Encourage decentralized experimentation and startup formation to explore uncharted technological frontiers.
- Preserve institutional flexibility:
  - Avoid overreliance on crude performance indicators (e.g., patent counts) that encourage low-value filings and gaming.
  - Design policies that reward competence and experimentation rather than patronage and loyalty.
- Safeguard legal protections for innovators:
  - Ensure legal frameworks protect firms from sudden political shifts that could divert resources toward building political alliances instead of innovation.

*Content adapted from "How the Battle for Control Could Crush AI’s Promise" by Carl Benedikt Frey, F&D Magazine, September 2025.*

---

## Content in this bundle

- **How the Battle for Control Could Crush AI's Promise**
  - [How the Battle for Control Could Crush AI's Promise (Markdown version)](/-/media/files/publications/fandd/article/2025/09/frey.pdf.md){rel="alternate" type="text/markdown"}
  - [How the Battle for Control Could Crush AI's Promise (PDF)](/-/media/files/publications/fandd/article/2025/09/frey.pdf){rel="external" type="application/pdf"}

---

_Source: https://www.imf.org/en/publications/fandd/issues/2025/09/how-the-battle-for-control-could-crush-ais-promise-carl-benedikt-frey_